Panorama Creation, Homography Estimation, and Feature Matching

Module: Computer Vision | Difficulty: Premium

Homography Matrix

Maps points between two planes:

with 8 degrees of freedom (up to scale).

Direct Linear Transform (DLT)

For each correspondence :

RANSAC Algorithm

Repeat times:

1. Sample 4 random correspondences
2. Compute via DLT
3. Count inliers where
4. Keep best model

where = confidence, = inlier ratio, = 4.

Multi-Band Blending (Laplacian Pyramid)

import cv2
import numpy as np

def find_matches(img1, img2, n_features=5000):
    sift = cv2.SIFT_create(nfeatures=n_features)
    kp1, des1 = sift.detectAndCompute(
        cv2.cvtColor(img1, cv2.COLOR_BGR2GRAY), None)
    kp2, des2 = sift.detectAndCompute(
        cv2.cvtColor(img2, cv2.COLOR_BGR2GRAY), None)
    bf = cv2.BFMatcher()
    matches = bf.knnMatch(des1, des2, k=2)
    good = [m for m, n in matches if m.distance < 0.75 * n.distance]
    pts1 = np.float32([kp1[m.queryIdx].pt for m in good])
    pts2 = np.float32([kp2[m.trainIdx].pt for m in good])
    return pts1, pts2, good

def compute_homography_ransac(pts1, pts2, threshold=5.0, max_iters=2000):
    best_H = None
    best_inliers = 0
    n = len(pts1)
    for _ in range(max_iters):
        idx = np.random.choice(n, 4, replace=False)
        H, _ = cv2.findHomography(pts1[idx], pts2[idx])
        if H is None:
            continue
        pts1_h = np.hstack([pts1, np.ones((n, 1))])
        projected = (H @ pts1_h.T).T
        projected = projected[:, :2] / projected[:, 2:3]
        errors = np.linalg.norm(projected - pts2, axis=1)
        inliers = (errors < threshold).sum()
        if inliers > best_inliers:
            best_inliers = inliers
            best_H = H
    return best_H

def stitch_images(img1, img2, H):
    h1, w1 = img1.shape[:2]
    h2, w2 = img2.shape[:2]
    result = cv2.warpPerspective(img2, H, (w1 + w2, max(h1, h2)))
    result[:h1, :w1] = img1
    return result

Research Insight: Traditional stitching pipelines (detect, match, estimate, blend) are being replaced by learned approaches like StitchFormer, which jointly optimize feature extraction and alignment. The key challenge in multi-image stitching is drift accumulation from sequential pairwise homography estimation. Global bundle adjustment minimizes the reprojection error across all images simultaneously, producing geometrically consistent panoramas. Multi-band blending avoids seams by fusing images at different spatial frequencies.